⭐ Star VAE LTXV Save

LTX Video checkpoints are all-in-one files: model, text encoder, and two VAEs (one for video, one for audio) packed into a single safetensors. That's convenient until you want to use those VAEs somewhere the whole checkpoint isn't welcome - like a hand-built graph that loads a video VAE from models/vae. Star VAE LTXV Save is a one-time setup tool that cracks the checkpoint open and writes the two VAEs out as standalone files you can load anywhere.

How it works

Point it at an LTX checkpoint and it does the split for you:

• Select checkpoint from your models/checkpoints folder.
• Give it a save_name (default ltxv_vae).

On run, it reads the checkpoint's state dict and separates the tensors by key prefix: everything under audio_vae.* becomes the audio VAE, and everything under vae.* becomes the video VAE (with the prefix stripped). The results are written to your models/vae folder as:

• <save_name>_video.safetensors
• <save_name>_audio.safetensors

The info output (STRING) tells you exactly what it wrote and how many tensors each file got - useful for sanity-checking that the split actually captured both halves. This is the "make the files exist" step that Star VAE LTXV Load depends on.

Why it's worth doing

Two reasons. First, the LTX all-in-one nodes load everything internally, but if you want to mix an LTX VAE with a different model - say, decoding LTX video latents with a standalone VAE in a custom pipeline - you need the loose files. Second, it's a disk-space and memory win: loading a 2B+ checkpoint just to grab the VAE is wasteful; with the split files, a VAE-only load is tiny and fast. It also future-proofs your workflow against checkpoint churn - split once, keep the small VAEs, swap checkpoints freely.

Installing

Part of StarNodes:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
cd ComfyUI_StarNodes
pip install -r requirements.txt

Restart, search star under ⭐StarNodes/Video. ComfyUI Manager users: search "Starnodes".

Gotchas

The obvious one: it only works on checkpoints that actually contain both vae.* and audio_vae.* keys. A plain SD1.5 or SDXL checkpoint has neither, and the node will happily write a video VAE file with (possibly) some content and an audio VAE with zero tensors - read the info output and don't skip it. Also note it writes into your models/vae folder, not the output folder, so you won't find the results where save nodes normally put things. Run it once, keep the files, and never think about the split again.